(710d) Network Insights into Improving Drug Target Prediction Accuracy of Inference Algorithms

Understanding disease mechanisms or the mechanism of action (MoA) of drugs requires systematic modeling of biological data. This is because a disease or chemical compound can trigger a variety of signaling pathways, which results in complex dynamic changes in a biological system. Protein target or equivalently drug target inference algorithms predict the molecular targets that are perturbed by a disease or a chemical compound by combining network characteristics and gene expression. Showing promising results in recent studies [1-3], these algorithms have provided a novel approach in detecting key factors driving disease outcomes, characterizing drug targets, and identifying drug toxicity.

However, as protein target inference algorithms become more complicated, it is unclear what roles gene expression and network data play in predicting these targets. It remains an open question which kind of biological data among protein-protein interactions (PPIs), protein-gene interactions (PGIs) and gene expression profiles most effects the accuracy of inferring disease/drug targets. This prevents researchers from improving mathematical methods that put more weight on the effective type of data to improve inference of targets.

Here, we evaluate how the selection of PPIs and the network characteristics of proteins within the select PPIs impact target inference using a standardized dataset from the DREAM challenge, which contains microarray data of human B cells treated with drugs having known targets [4]. DeMAND [1] and ProTINA [2] serve as representative inference algorithms which have already been tested on this dataset. In addition, we predict protein targets by applying topological analysis to PPI and PGI networks. Both the targets predicted by the two algorithms and those obtained from topological analysis are compared with the known targets of associated drugs. We find that topological analysis has high accuracy in predicting drug targets and the prediction performance is comparable to that of both of the more advanced algorithms. Moreover, combining gene expression with network degree into an overall score for the likelihood of being the drug target outperforms both algorithms. Overall, the analyses indicate that network topology, and not gene expression or even dynamic gene expression, drives accurate discovery of drug targets. Future work on inferring protein targets for understanding disease mechanisms and the MoA of drugs should focus on improving the consensus for confident PPI and PGI networks. And future algorithms should more carefully weight highly interactive genes.

1. Woo, J.H., et al., Elucidating Compound Mechanism of Action by Network Perturbation Analysis. Cell, 2015. 162(2): p. 441-451.
2. Noh, H., J.E. Shoemaker, and R. Gunawan, Network perturbation analysis of gene transcriptional profiles reveals protein targets and mechanism of action of drugs and influenza A viral infection. Nucleic Acids Res, 2018.
3. Jain, S., et al., Reconstructing the temporal progression of HIV-1 immune response pathways. Bioinformatics, 2016. 32(12): p. i253-i261.
4. Bansal, M., et al., A community computational challenge to predict the activity of pairs of compounds. Nat Biotechnol, 2014. 32(12): p. 1213-22.